Classifying Gait Alterations Using an Instrumented Smart Sock and Deep Learning
Classifying Gait Alterations Using an Instrumented Smart Sock and Deep Learning
复制标题
DOI：
10.1109/jsen.2022.3216459
复制
发表时间：
2022-12-01
影响因子：
4.3
通讯作者：
Dias, Tilak
中科院分区：
文献类型：
--
作者：
Lugoda, Pasindu;Hayes, Stephen Clive;Dias, Tilak
本文提出了一种非侵入性方法，利用不显眼的仪器化袜子和深度学习网络，对与各种运动障碍和/或神经系统疾病相关的步态模式进行分类。无缝仪器袜是使用三根嵌入加速度计的纱线制成的，分别位于脚趾（拇趾）、脚跟上方和外踝上。对 12 名身体健全的参与者进行了人体试验，每只脚都穿了一只装有仪器的袜子。参与者被要求完成七项试验，包括他们的典型步态和六种不同的步态类型，模仿与各种运动障碍和神经系统疾病相关的典型运动模式。对四个神经网络和一个支持向量机进行了测试，以确定自动数据分类的最有效方法。双长短期记忆 (LSTM) 生成了最准确的结果，并表明与每英尺使用单个加速度计相比，每英尺使用三个加速度计的分类精度提高了 11.4%。当仅使用单个加速度计进行分类时，与其他两个加速度计相比，脚踝加速度计生成的结果最准确。该网络能够正确分类五种不同的步态类型：跺脚（100%）、慢步（66.8%）、双瘫（66.6%）、偏瘫（66.6%）和“正常行走”（58.0%）。该网络无法正确区分足拍步态 (21.2%) 和踏步步态 (4.8%)。这项工作表明，包含三种加速度计纱线的仪器化纺织袜能够生成足够的数据，使神经网络能够区分特定的步态模式。这可以使临床医生和治疗师能够对步态改变进行远程分类并观察康复期间步态的变化。
This article presents a noninvasive method of classifying gait patterns associated with various movement disorders and/or neurological conditions, utilizing unobtrusive, instrumented socks and a deep-learning network. Seamless instrumented socks were fabricated using three accelerometer-embedded yarns, positioned at the toe (hallux), above the heel, and on the lateral malleolus. Human trials were conducted on 12 able-bodied participants, an instrumented sock was worn on each foot. Participants were asked to complete seven trials consisting of their typical gait and six different gait types that mimicked the typical movement patterns associated with various movement disorders and neurological conditions. Four neural networks and an SVM were tested to ascertain the most effective method of automatic data classification. The bi-long short-term memory (LSTM) generated the most accurate results and illustrates that the use of three accelerometers per foot increased classification accuracy compared to a single accelerometer per foot by 11.4%. When only a single accelerometer was utilized for classification, the ankle accelerometer generated the most accurate results in comparison to the other two. The network was able to correctly classify five different gait types: stomp (100%), shuffle (66.8%), diplegic (66.6%), hemiplegic (66.6%), and "normal walking " (58.0%). The network was incapable of correctly differentiating foot slap (21.2%) and steppage gait (4.8%). This work demonstrates that instrumented textile socks incorporating three accelerometer yarns were capable of generating sufficient data to allow a neural network to distinguish between specific gait patterns. This may enable clinicians and therapists to remotely classify gait alterations and observe changes in gait during rehabilitation.